---
license: apache-2.0
language:
  - en
  - zh
tags:
  - robotics
  - grasping
  - vision-language-model
  - locate-anything
  - grasp-rect
pipeline_tag: image-text-to-text
---

# Grasp Anything 2D — Checkpoint 9040 (grasp_rect 专用)

基于 NVIDIA [LocateAnything-3B](https://huggingface.co/nvidia/LocateAnything-3B) 的语言引导二维抓取模型,Phase 2 SFT 阶段产物。

## 任务

`grasp_rect` — 输出 `(cx, cy, θ, width)` 矩形抓取参数,按 RealVLG 固定 40 px depth 展开矩形。

## 评测结果(RealVLG 官方 652 样本)

| Split | n | gAcc_corrected_strict | mIoU_strict | angle_seam_gAcc |
|---|---|---|---|---|
| seen | 253 | **69.96%** | 43.08% | 65.00% |
| similar | 235 | 57.02% | 35.48% | 81.25% |
| novel | 164 | 27.44% | 17.02% | 33.33% |

- 评测协议: `evaluate_realvlg_grasp.py`,fast 模式
- gAcc_corrected_strict: IoU>0.25 且 angle<30° (mod 180, 修正了官方 radian/degree bug)
- 对比 RealVLG-R1 GRPO Grasp 3B: seen gAcc +28.5%, similar +22.1%

## 使用方法

```python
from locate_anything_service.model import LocateAnythingRuntime
from locate_anything_service.config import Settings

settings = Settings(model_id="charlesH777/grasp-anything-9040")
runtime = LocateAnythingRuntime(settings)
result = runtime.predict("image.jpg", "抓取红色杯子", mode="grasp_rect")
```

## 训练配置

- 阶段: Phase 2 SFT (grasp_rect 专用)
- LLM LoRA: rank 32, target Qwen2.5-3B-Instruct
- 训练参数: `grasp_rect_task_enabled: true`, `contact_loss_enabled: false`